[MachineVerifier] Doing ::calcRegsPassed in RPO: ~35% faster MV

[rtereshin]

Summary

Depending on the target, test suite, pipeline config and perhaps other
factors machine verifier when forced on with -verify-machineinstrs can
increase compile time 2-2.5 times over (Release, Asserts On), taking up
~60% of the time. An invaluable tool, it significantly slows down
machine verifier-enabled testing.

Nearly 75% of its time MachineVerifier spends in the calcRegsPassed
method. It's a classic forward dataflow analysis executed over sets, but
visiting MBBs in arbitrary order. We switch that to RPO here.

This speeds up MachineVerifier by about 35%, decreasing the overall
compile time with -verify-machineinstrs by 20-25% or so.

calcRegsPassed itself gets 2x faster here.

All measured on a large suite of shaders targeting a number of GPUs.

[bogner]

llvm/lib/CodeGen/MachineVerifier.cpp

2133–2134

Not sure it's worth explaining this tradeoff in a comment - the explanation above is good for what we're doing here and the commit message is pretty clear on why. Feel free to leave as is if you think it's valuable though.

[rtereshin]

llvm/lib/CodeGen/MachineVerifier.cpp

2133–2134

Nah, I agree, trimming the comment (as further briefly discussed offline).

Thanks!

[rtereshin]

Trimming down the comments a little bit as requested.

llvm/lib/CodeGen/MachineVerifier.cpp

 // The machine code verifier is enabled with the command-line option
 // -verify-machineinstrs.
 //===----------------------------------------------------------------------===//
 
 #include "llvm/ADT/BitVector.h"
 #include "llvm/ADT/DenseMap.h"
 #include "llvm/ADT/DenseSet.h"
 #include "llvm/ADT/DepthFirstIterator.h"
+#include "llvm/ADT/PostOrderIterator.h"
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/SetOperations.h"
 #include "llvm/ADT/SmallPtrSet.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/StringRef.h"
 #include "llvm/ADT/Twine.h"
 #include "llvm/Analysis/EHPersonalities.h"
 #include "llvm/CodeGen/GlobalISel/RegisterBank.h"
if (Indexes) {
     lastIndex = stop;
   }
 }
 
 // Calculate the largest possible vregsPassed sets. These are the registers that
 // can pass through an MBB live, but may not be live every time. It is assumed
 // that all vregsPassed sets are empty before the call.
 void MachineVerifier::calcRegsPassed() {
+  // This is a forward dataflow, doing it in RPO. A standard map serves as a
+  // priority (sorting by RPO number) queue, deduplicating worklist, and an RPO
+  // number to MBB mapping all at once.
+  std::map<unsigned, const MachineBasicBlock *> RPOWorklist;
+  DenseMap<const MachineBasicBlock *, unsigned> RPONumbers;
+  if (MF->empty()) {
+    // ReversePostOrderTraversal doesn't handle empty functions.
+    return;
+  }
+  for (const MachineBasicBlock *MBB :
+       ReversePostOrderTraversal<const MachineFunction *>(MF)) {
+    // Careful with the evaluation order, fetch next number before allocating.
+    unsigned Number = RPONumbers.size();
+    RPONumbers[MBB] = Number;
+  }
   // First push live-out regs to successors' vregsPassed. Remember the MBBs that
   // have any vregsPassed.
-  SmallPtrSet<const MachineBasicBlock*, 8> todo;
-  for (const auto &MBB : *MF) {
+  for (const MachineBasicBlock &MBB : *MF) {
     BBInfo &MInfo = MBBInfoMap[&MBB];
     if (!MInfo.reachable)
       continue;
-    for (MachineBasicBlock::const_succ_iterator SuI = MBB.succ_begin(),
-           SuE = MBB.succ_end(); SuI != SuE; ++SuI) {
-      BBInfo &SInfo = MBBInfoMap[*SuI];
+    for (const MachineBasicBlock *Succ : MBB.successors()) {
+      BBInfo &SInfo = MBBInfoMap[Succ];
       if (SInfo.addPassed(MInfo.regsLiveOut))
-        todo.insert(*SuI);
+        RPOWorklist.emplace(RPONumbers[Succ], Succ);
     }
   }
 
-  // Iteratively push vregsPassed to successors. This will converge to the same
-  // final state regardless of DenseSet iteration order.
-  while (!todo.empty()) {
-    const MachineBasicBlock *MBB = *todo.begin();
-    todo.erase(MBB);
+  // Iteratively push vregsPassed to successors.
+  while (!RPOWorklist.empty()) {
+    auto Next = RPOWorklist.begin();
+    const MachineBasicBlock *MBB = Next->second;
+    RPOWorklist.erase(Next);
     BBInfo &MInfo = MBBInfoMap[MBB];
-    for (MachineBasicBlock::const_succ_iterator SuI = MBB->succ_begin(),
-           SuE = MBB->succ_end(); SuI != SuE; ++SuI) {
-      if (*SuI == MBB)
+    for (const MachineBasicBlock *Succ : MBB->successors()) {
+      if (Succ == MBB)
         continue;
-      BBInfo &SInfo = MBBInfoMap[*SuI];
+      BBInfo &SInfo = MBBInfoMap[Succ];
       if (SInfo.addPassed(MInfo.vregsPassed))
-        todo.insert(*SuI);
+        RPOWorklist.emplace(RPONumbers[Succ], Succ);
     }
   }
 }
 
 // Calculate the set of virtual registers that must be passed through each basic
 // block in order to satisfy the requirements of successor blocks. This is very
 // similar to calcRegsPassed, only backwards.
 void MachineVerifier::calcRegsRequired() {